We conducted a comprehensive search of PubMed, Embase, and Web of Science databases from when the database was created until December 1, 2022. The search keywords include “immune checkpoint inhibitors”, “immune checkpoint inhibitor-related adverse reaction”, “risk factor”, “blood biomarker”, “eosinophil”, “neutrophil”, “lymphocyte”, “monocyte”, etc. The specific search formula can be found in the appendix.

The inclusion criteria were as follows (1): Patients had malignant tumors (2); The study was a randomized clinical trial, retrospective clinical study, or case-control study; (3) Patients received ICI therapy; (4) The association between blood cell counts and the occurrence of irAEs was assessed by the authors; and (5) Definite odds ratio (OR) and 95% CI. If no OR value is specified, at least specific values that can be calculated should be provided.

The exclusion criteria were as follows: (1) The study type was animal experiments; (2) Risk factors for the occurrence of irAEs were not assessed; (3) Blood counts were not considered to be risk factors; and (4) Patients had hematologic tumors.

After an initial screening based on the title and abstract of the publication and excluding any irrelevant literature, the remaining literature was read in total, and those that should be included were selected.

Two independent investigators identified the included papers, and a third investigator resolved disagreements, if any. The following information was extracted: title, author, publication year, country, patient number, patient origin, research type, risk factors, type of irAEs, optimal cutoff value, OR, and 95% CI. If univariate and multivariate regression analyses were conducted, ORs were calculated using the multivariate regression analysis data. When OR values were unavailable, crude OR values and 95% CIs were calculated based on frequency tables.

The quality in prognosis studies (QUIPS) tool was used to evaluate the risk of bias. Briefly, two independent researchers divided the literature into low, moderate, and high risk-of-bias categories. Any disagreement was discussed by adding a third person until an agreement was reached.

The same risk factor was analyzed using the “metan” command in Stata17. Data included crude or adjusted ORs and 95% CIs, with ORs adjusted for confounders selected whenever possible. Additionally, the heterogeneity test was conducted, and I² > 50% was considered highly heterogeneous.

The three databases were searched with the set search formula, and 851 publications fit the criteria. Among them, 188 meta-analyses and systematic reviews were excluded. After reading the title and abstract, 584 publications were excluded. Further filtering was done by reading the full text. Finally, 18 studies were included. Details can be found in the flow chart. ( Figure 1 )

A total of 18 studies (11, 15–31) were included in the analysis. These studies were published between 2019 to 2022. Nine studies took place in China, six in Japan, one in Singapore, one in Italy, one in Canada, which shows that most of these studies were involved the Asian population. A total of 3579 patients were included in the study. Previous treatment was described in four articles (15–18), one of which included previous antitumor regimens as one of the criteria when including patients (16). Six articles included only patients who received ICI monotherapy (15, 16, 19–22). The remaining studies were unrestricted, and they included patients who received ICI in combination with other antitumor treatments. With immune checkpoint inhibitor administration as the node, blood cell counts from all studies were evaluated. Data from either the beginning or end of immunotherapy were chosen for analysis. The types of blood counts reported are as follows: absolute eosinophil count (AEC), derived neutrophil: lymphocyte ratio (dNLR), neutrophil: lymphocyte ratio (NLR), absolute neutrophil count (ANC), absolute lymphocyte count (ALC), absolute monocyte count (AMC), platelet count (PLT), monocyte: lymphocyte ratio (MLR), platelet: lymphocyte ratio (PLR), etc. Eight studies showed substantial truncation values (11, 15–17, 20, 23, 25, 28). Eleven studies reported the use of receiver operating characteristic (ROC) curve analysis to calculate the optimal cutoff value (11, 15–17, 19–21, 23–26). In all included reports, skin adverse reactions were the most common (23.34%), followed by endocrine adverse reactions (22.81%), immune pneumonia (13.50%), hepatic adverse reactions (10.25%), and gastrointestinal adverse reactions (9.23%). All studies were retrospective. More details are shown in Supplementary Table 1 .

QUIPS tool was used to assess the risk of bias; detailed entries include study participation, study attrition, prognostic factor measurement, outcome measurement, study confounding, statistical analysis and reporting. After assessing the articles, seven were classified as high risk of bias, six as medium risk, and five as low risk. ( Supplementary Figure 1 )

The same risk factor reported in at least two publications was included in the study, with preference given to data reporting cutoff values. Further, the results of multifactorial analyses were prioritized for inclusion, and if unavailable, single-factor results were considered for inclusion. Available for analysis are AEC, NLR, and PLR.

Six papers described AEC, three of which indicated a cut-off value. We analyzed articles that provided a cut-off value separately and those that did not. The meta-analysis, including only cut-off values, yielded an OR of 2.09 (95% CI: 1.28–3.43). Studies that did not report cut-off values had an OR of 1.01 (95% CI: 1.00–1.02). ( Figure 2 ) In this analysis, 1202 cases were involved, the most frequent being lung cancer (70.97%), followed by renal and urological tumors (10.73%), malignant melanoma (4.99%), and head and neck tumors (4.41%).

The effect of pre-treatment NLR on the incidence of irAEs was mainly described in studies reporting NLR; two articles reported the effect of post-treatment values. We divided the analysis into three parts: baseline with and without cutoff values and post-treatment data. The most statistically significant studies were those accompanied by a truncated value at baseline, with an OR of 2.28 (95%CI: 1.49–3.48). ( Figure 3 ) The studies that provided cut-off values included a total of 1116 cases, with lung cancer (87.81%), renal and urinary tumors (5.29%), head and neck tumors (1.52%), and esophageal cancer (1.52%).

Among the articles presenting PLR, one presenting post-treatment values was not analyzed. The study that reported a cutoff value was statistically significant with an OR of 1.70 (95% CI: 1.04-2.78). ( Figure 4 ) A total of 907 cases were involved in the studies that reported cut-off values, including lung cancer (86.44%), renal and urological tumors (6.28%), esophageal cancer (1.87%), and liver cancer (1.21%).

We present data from studies that reported cut-off values and analyzed them further. ( Table 1 ) Studies involving AEC have concluded that higher than optimal threshold values are associated with the occurrence of adverse reactions in patients. Of the studies that reported NLR cutoff values, only one concluded that being above the cutoff value caused an increased incidence of adverse reactions (20), whereas the remaining studies concluded that being below the cutoff value was associated with an event rate of a combined OR 2.52 (95%CI:1.49-4.52). While in PLR-related studies, PLR above the cutoff (OR: 1.43, 95% CI: 0.75–2.75) was analyzed, but considering that one of the articles did not show a correlation and the comparison of OR values, we think the below the cutoff has a higher correlation with event occurrence. Further analysis of the specific range of values chosen can be seen in the subgroup analysis.

Fewer articles presented truncation values of ANC, ALC, AMC, PLT, and LMR. Therefore, we did not perform further analysis.

Heterogeneity was observed in the meta-analysis of all three indicators. Funnel plots indicate possible publication bias, and studies with positive results are more likely to be selected. (Figure not shown)

Sensitivity analyses were performed by sequentially eliminating individual studies to clarify whether each study affected the overall results. The analysis gave stable results for sensitivity analysis in the NLR and PLR studies that reported cutoff values and in the AEC studies that did not report cutoff values. In contrast, sensitivity analysis showed that the results might be unstable and not statistically significant in other studies. (Figure not shown)

Subgroup analyses were performed for PLR and NLR based on country, the number of cases, and cutoff values. Due to the limited number of studies, the remaining risk factors were not included in these additional analyses. The country-based subgroups showed that NLR was more significant in the incidence of irAEs in Japanese patients with an OR of 3.40 (95% CI: 1.82-6.32), significantly higher than that in the Chinese population had an OR of 0.96 (95%CI: 0.38-2.40). The number of studies that did not report NLR cutoff values was insufficient for subgroup analysis. PLR contributed more significantly to the incidence of irAEs in the Japanese population than in the Chinese population, with an OR of 3.78 (95% CI:1.84-7.78) vs 0.86 (95% CI:0.44-1.67), respectively. We further performed a stratified analysis of the number of patients included in the study. Statistical significance was most significant when the number of people included in studies on NLR was less than 100. Effect values did not differ significantly between patient numbers in the PLR group where cutoff values were reported. In addition, four articles reported NLR cutoff values between 2.5 and 3.5, and their OR was calculated to be 2.13 (95%CI: 1.34-3.40). Due to the limited number of studies, there may be inaccuracies in the analysis results for each subgroup. All the above results are represented in Supplementary Table 2 .